Electricity Demand Forecasting for the Luzon Power System

Data Scientist Pathway

Abstract

The increasing electricity demand requires all power system stakeholders to plan their operations, maintenance schedules, and capacity expansions of infrastructures, among others, with utmost accuracy to provide high-quality service to their customers. Specifically, different types of power system planning provide various benefits: operational planning ensures the economical dispatch of generators and low-cost trading of electricity; maintenance planning determines the optimal schedule in shutting down generators for maintenance such that the remaining online generators could still meet the demand and no widespread power interruption is triggered; and infrastructure planning assesses the right capacity of generators, power lines, and other electrical equipment to be installed so as not to have an under-designed or over-designed system. Such planning endeavors start with evaluating the needs of consumers by electricity demand profiling and subsequently, electricity demand forecasting.

In Luzon, numerous yellow alerts (i.e., reserve is insufficient to cover the largest running generating unit) and red alerts (i.e., reserve is zero, a generation deficiency exists, or there is critical loading or imminent overloading of lines) have been recorded in the past years. The Department of Energy (DOE) further estimates at least 12 yellow alerts this 2023, with unfinished power projects and unforeseen forced outages and maintenance of power plants being among the main causes of low power supply. In order to avoid these alerts, it is imperative to perform adequate power system planning that is rooted from electricity demand forecasting. This study serves as an example of how electricity demand forecasting can be done for the Luzon power system by using various forecasting techniques and models, namely, (1) a quadratic trend model, (2) a quadratic trend model with seasonality effects, and (3) a neural network with tuned number of neurons per layer.

Monthly electricity peak demand data from 2001 to 2015 and 2016 to 2020 were used to train and test the models, respectively. The three models were generated using different tools of Excel and Python: model (1) was built using the Regression tool of Excel; model (2) was created using the Solver tool of Excel; and model (3) was trained using Python with the Keras library. All models yielded a relatively good fit with respect to the training set (2001 to 2015) with model (2) having the lowest RMSE at 155.9510 MW and MAPE at 1.8762%. Moreover, all models had a relatively low error when tested using the test set (2016 to 2020) with model (3) having the lowest MAPE at 5.1601% and model (2) having the lowest RMSE at 629.4646 MW. It should be emphasized, however, that the lowest RMSE for the test set of 629.4646 MW represents an estimate of the tolerance level of the forecast, hence, this tolerance may or may not be endurable depending on the application. If this tolerance is considered in an operational level, then a 629.4646 MW discrepancy in planning may cause widespread power interruptions, but this value could be tolerable for maintenance and infrastructure planning applications.